Closed-Loop, Multiobjective Optimization of Two-Dimensional Gas

Metabolomics seeks to measure potentially all the metabolites in a biological sample, and consequently, we need to develop and optimize methods to inc...
0 downloads 0 Views 929KB Size
Download PDF
Anal. Chem. 2007, 79, 464-476

Closed-Loop, Multiobjective Optimization of Two-Dimensional Gas Chromatography/Mass Spectrometry for Serum Metabolomics Steve O’ Hagan, Warwick B. Dunn, Joshua D. Knowles, David Broadhurst, Rebecca Williams, Jason J. Ashworth, Maureen Cameron, and Douglas B. Kell*

School of Chemistry, The University of Manchester, Faraday Building, Sackville Street, Manchester M60 1QD, UK, and The Manchester Interdisciplinary Biocentre, The University of Manchester, 131 Princess Street, Manchester M1 7DN, UK

Metabolomics seeks to measure potentially all the metabolites in a biological sample, and consequently, we need to develop and optimize methods to increase significantly the number of metabolites we can detect. We extended the closed-loop (iterative, automated) optimization system that we had previously developed for onedimensional GC-TOF-MS (O’Hagan, S.; Dunn, W. B.; Brown, M.; Knowles, J. D.; Kell, D. B. Anal. Chem. 2005, 77, 290-303) to comprehensive two-dimensional (GC×GC) chromatography. The heuristic approach used was a multiobjective version of the efficient global optimization algorithm. In just 300 automated runs, we improved the number of metabolites observable relative to those in 1D GC by some 3-fold. The optimized conditions allowed for the detection of over 4000 raw peaks, of which some 1800 were considered to be real metabolite peaks and not impurities or peaks with a signal/noise ratio of less than 5. A variety of computational methods served to explain the basis for the improvement. This closed-loop optimization strategy is a generic and powerful approach for the optimization of any analytical instrumentation. There is increasing interest in the measurement of nominally “all” the metabolites in a sample, i.e., the metabolome.1-5 In practice, the very wide chemical and physical nature of these metabolites6,7 means that only a subset, a metabolic profile, is determined using a given technique.8,9 Nevertheless, from the philosophical point of view, in which we use these methods * To whom correspondence should be addressed. Phone: 0044 161 306 4492. E-mail: [email protected]; www.dbkgroup.org. (1) Oliver, S. G.; Winson, M. K.; Kell, D. B.; Baganz, F. Trends Biotechnol. 1998, 16, 373-378. (2) Harrigan, G. G., Goodacre, R., Eds. Metabolic profiling: its role in biomarker discovery and gene function analysis; Kluwer Academic Publishers: Boston, 2003. (3) Tomita, M., Nishioka, T., Eds. Metabolomics: the frontier of systems biology; Springer: Tokyo, 2005. (4) Vaidyanathan, S., Harrigan, G. G., Goodacre, R., Eds. Metabolome analyses: strategies for systems biology; Springer: New York, 2005. (5) van der Greef, J.; Stroobant, P.; van der Heijden, R. Curr. Opin. Chem. Biol. 2004, 8, 559-565. (6) Nobeli, I.; Ponstingl, H.; Krissinel, E. B.; Thornton, J. M. J. Mol. Biol. 2003, 334, 697-719. (7) Nobeli, I.; Thornton, J. M. Bioessays 2006, 28, 534-545.

464 Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

principally for hypothesis generation rather than hypothesis testing,10 we do seek methods that can measure as many of the metabolites as possible to maximize the biological information obtained. Of methods currently in use,9,11,12 those that employ a separation step coupled to mass spectrometric detection are pre-eminent. While advances in capillary electrophoresis13,14 and LC15,16 are making them more technically competitive, both pioneering (e.g., refs 17-20) and more recent studies (e.g., refs 21 and 22) have favored gas chromatography as being the most highly resolving technique, and thus the separation method of choice. Despite this long history, gas chromatographic separations have been far from optimized. The reason for this is that a comparatively large set (m) of instrumental parameters may be varied, and the number of combinations varies exponentially with m such that if each can take n values the number of possible experiments is nm. For even modest values of n and m, exhaustive search becomes impossible. Notwithstanding, it has long been known that comparatively small changes in experimental conditions can have rather substantial effects on chromatographic performance, especially in liquid chromatography.23-25 The same is also true for electrospray ionization mass spectrometry (e.g., (8) Fiehn, O. Comp. Funct. Genomics 2001, 2, 155-168. (9) Goodacre, R.; Vaidyanathan, S.; Dunn, W. B.; Harrigan, G. G.; Kell, D. B. Trends Biotechnol. 2004, 22, 245-252. (10) Kell, D. B.; Oliver, S. G. Bioessays 2004, 26, 99-105. (11) Dunn, W. B.; Ellis, D. I. Trends Anal. Chem. 2005, 24, 285-294. (12) Dunn, W. B.; Bailey, N. J. C.; Johnson, H. E. Analyst 2005, 130, 606-625. (13) Soga, T.; Ueno, Y.; Naraoka, H.; Ohashi, Y.; Tomita, M.; Nishioka, T. Anal. Chem. 2002, 74, 2233-2239. (14) Soga, T.; Ohashi, Y.; Ueno, Y.; Naraoka, H.; Tomita, M.; Nishioka, T. J. Proteome Res. 2003, 2, 488-494. (15) Plumb, R.; Castro-Perez, J.; Granger, J.; Beattie, I.; Joncour, K.; Wright, A. Rapid Commun. Mass Spectrom. 2004, 18, 2331-2337. (16) Wilson, I. D.; Nicholson, J. K.; Castro-Perez, J.; Granger, J. H.; Johnson, K. A.; Smith, B. W.; Plumb, R. S. J. Proteome Res. 2005, 4, 591-598. (17) Horning, E. C.; Horning, M. G. Clin. Chem. 1971, 17, 802-809. (18) Tanaka, K.; West-Dull, A.; Hine, D. G.; Lynn, T. B.; Lowe, T. Clin. Chem. 1980, 26, 1847-1853. (19) Jellum, E.; Bjornson, I.; Nesbakken, R.; Johansson, E.; Wold, S. J. Chromatogr. 1981, 217, 231-237. (20) Greenaway, W.; May, J.; Scaysbrook, T.; Whatley, F. R. Z. Naturforsch., C 1991, 46, 111-121. (21) Fiehn, O.; Kopka, J.; Dormann, P.; Altmann, T.; Trethewey, R. N.; Willmitzer, L. Nat. Biotechnol. 2000, 18, 1157-1161. 10.1021/ac061443+ CCC: $37.00

© 2007 American Chemical Society Published on Web 12/15/2006

refs 26-28). Another issue, of course, is that we do not know the numbers of metabolites that might be present in a given matrix, although for serum we would argue that values for the number of native metabolites in the decade 1-10 000 seem reasonable based on a combination of our knowledge of the major metabolic pathways (e.g., refs 29-32) and what has been observed in the more modern experiments designed to explore this question (e.g., ref 33). We note too the potential contribution of the gut microflora to the serum metabolome,34 which may be more pronounced in urine.16 In recent work,35 inspired by the “Robot Scientist” idea,36 we exploited a closed-loop method in which we automated and iterated the entire process of parameter setting, performance of the GC run, analysis of the data obtained (in terms of peak number, run time, and a metric of signal/noise ratio), and changing of the parameter set, with the result that with just 240 runs we could improve an already excellent method 3-fold in terms of the number of peaks detected. A multiobjective evolutionary algorithm, PESA-II ,35,37,38 was used as the heuristic for navigating the search space of some 200 000 000 combinations. Comprehensivetwo-dimensionalgaschromatography(GC×GC)39-44 describes a general method in which substances eluting from a first column (typically nonpolar) over a certain time window are focused and then released (via a process termed modulation) on to a second (typically more polar) column where they are further separated. The technique has the potential for increasing further (22) Catchpole, G. S.; Beckmann, M.; Enot, D. P.; Mondhe, M.; Zywicki, B.; Taylor, J.; Hardy, N.; Smith, A.; King, R. D.; Kell, D. B.; Fiehn, O.; Draper, J. Proc. Natl. Acad. Sci. U.S.A. 2005, 102, 14458-14462. (23) Glajch, J. L.; Kirkland, J. J.; Snyder, L. R. J. Chromatogr. 1982, 238, 269280. (24) Kirkland, J. J.; Glajch, J. L. J. Chromatogr. 1983, 255, 27-39. (25) Glajch, J. L.; Kirkland, J. J.; Minor, J. M. J. Liq. Chromatogr. 1987, 10, 1727-1747. (26) Vaidyanathan, S.; Broadhurst, D. I.; Kell, D. B.; Goodacre, R. Anal. Chem. 2003, 75, 6679-6686. (27) Vaidyanathan, S.; Kell, D. B.; Goodacre, R. Anal. Chem. 2004, 76, 50245032. (28) Moberg, M.; Markides, K. E.; Bylund, D. J. Mass Spectrom. 2005, 40, 317324. (29) Goto, S.; Okuno, Y.; Hattori, M.; Nishioka, T.; Kanehisa, M. Nucleic Acids Res. 2002, 30, 402-404. (30) Duarte, N. C.; Herrgard, M. J.; Palsson, B. Ø. Genome Res. 2004, 14, 12981309. (31) Kanehisa, M.; Goto, S.; Kawashima, S.; Okuno, Y.; Hattori, M. Nucleic Acids Res. 2004, 32, D277-280. (32) Kanehisa, M.; Goto, S.; Hattori, M.; Aoki-Kinoshita, K. F.; Itoh, M.; Kawashima, S.; Katayama, T.; Araki, M.; Hirakawa, M. Nucleic Acids Res. 2006, 34, D354-357. (33) Want, E. J.; O’Maille, G.; Smith, C. A.; Brandon, T. R.; Uritboonthai, W.; Qin, C.; Trauger, S. A.; Siuzdak, G. Anal. Chem. 2006, 78, 743-752. (34) Nicholson, J. K.; Holmes, E.; Wilson, I. D. Nat. Rev. Microbiol. 2005, 3, 431-438. (35) O’Hagan, S.; Dunn, W. B.; Brown, M.; Knowles, J. D.; Kell, D. B. Anal. Chem. 2005, 77, 290-303. (36) King, R. D.; Whelan, K. E.; Jones, F. M.; Reiser, P. G. K.; Bryant, C. H.; Muggleton, S. H.; Kell, D. B.; Oliver, S. G. Nature 2004, 427, 247-252. (37) Corne, D.; Knowles, J.; Oates, M. Lecture Notes in Computer Science, Springer: Paris, France, 2000; pp 869-878. (38) Corne, D.; Jerram, N. R.; Knowles, J.; Oates, M., Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2001); Morgan Kaufmann: San Francisco, CA, 2001; pp 283-290. (39) Bertsch, W. HRC, J. High Resolut. Chromatogr. 2000, 23, 167-181. (40) Ong, R. C.; Marriott, P. J. J. Chromatogr. Sci. 2002, 40, 276-291. (41) Blumberg, L. M. J. Chromatogr., A 2003, 985, 29-38. (42) Marriott, P.; Shellie, R. Trends Anal. Chem. 2002, 21, 573-583. (43) van Mispelaar, V. G.; Tas, A. C.; Smilde, A. K.; Schoenmakers, P. J.; van Asten, A. C. J. Chromatogr., A 2003, 1019, 15-29. (44) Harynuk, J.; Marriott, P. J. Anal. Chem. 2006, 78, 2028-2034.

Figure 1. Closed-loop control of GC×GC-MS parameters using the ParEGO algorithm.

the number of peaks determined in metabolomics experiments caused by a combined effect of increased peak capacities, chromatographic resolution and signal-to-noise ratios.45 One distinct advantage is that metabolites of the same volatility, and hence the same retention time on column 1, can be resolved chromatographically on column 2 if their polarities are different, something that is not achievable with 1D GC and that, consequently, decreases the reliance on deconvolution software. With two columns, there are even more experimental parameters that can be varied, and correspondingly, it has not been subjected to extensive optimization, let alone the closed-loop optimization of the type described above. It was therefore of interest to develop a comprehensive two-dimensional GC×GC method that would maximize the number of metabolites we could observe, this time using human serum. This paper describes a successful implementation of closed-loop optimization in this system, leading to a method that produces more than 4000 peaks corresponding to almost 2000 different metabolite peaks. EXPERIMENTAL SECTION Biological Information. Sample Preparation. Deproteinization of human serum (pooled serum from 17 individuals therefore representing the optimal metabolome expected; SigmaAldrich, Gillingham, UK) was performed by addition of 600 µL of methanol (AR Grade, Sigma-Aldrich, Gillingham, UK) to 200 µL of serum followed by vortex mixing (15 s), centrifugation (15 min, 13385g), and lyophiliation of the supernatant (HETO VR MAXI vacuum centrifuge attached to a HETO CT/DW 60E cooling trap; Thermo Life Sciences, Basingstoke, UK). A two-stage chemical derivatization procedure was performed. A 50-µL aliquot of 20 mg/ mL O-methylhydroxylamine solution was added and heated at 40 °C for 80 min followed by addition of 50 µL of N-acetyl-N(trimethylsilyl)trifluoroacetamide and heating at 40 °C for 80 min. All sample solutions were analyzed within 36 h of derivatization. Optimization. The GC×GC-MS instrument (Agilent 6890N gas chromatograph (Agilent Technologies, Stockport, UK) and Gerstel MPS2L autosampler (Gerstel, Baltimore, MD) coupled (45) Welthagen, W.; Shellie, R. A.; Spranger, J.; Ristow, M.; Zimmermann, R.; Fiehn, O. Metabolomics 2005, 1, 65-73.

Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

465

Table 1. Instrument Parameters for Optimization Showing Final Ranges Useda variable parameters

units

min

sample volume corrected column flow split ratio inlet temperature oven 1 start hold time oven 1 ramp rate oven 1 final temperature oven 1 final hold time oven 2 start temperature transfer line temperature modulator temperature offset second dimension time hot pulse time acquisition rate ion source temperature

µL mL/min °C min °C/min °C min °C °C °C min sec Hz °C

1.0 0.8 1:10 200 3 5 260 0 55 220 15 4 0.2 30 220

fixed parameters oven 1 start temperature initial detector voltage

units °C V

default 50 1700

max

inc

step

5 2 1:80 280 8 26 300 5 75 280 55 7 0.5 160 280

1.0 (0.1) 0.2 (0.1) 5 (0.1) 10 (1) 1 (0.01) 3 (0.1) 10 (1) 1 (0.01) 4 (0.01) 20 (1) 10 (1) 1 (0.0001) 0.1 (0.01) 10 (1) 20 (1)

5 (45) 7 (12) 15 (700) 9 (40) 6 (500) 8 (95) 5 (40) 6 (400) 6 (2000) 4 (60) 5 (40) 4 (30000) 4 (30) 14 (440) 4 (60)

a Increments and steps for experiments 1-217 are shown wihout parentheses, and for experiments 218 and higher are shown in parentheses. Oven 2 start hold time, ramp speed, final temperature, and final temperature hold time were identical to oven 1 parameters. It will be noted that the final search space, which is the product of the numbers in the right-hand column of the variable parameters, was 7.32 × 109 for the initial search (and ∼9.8 × 1032 for the quasi-continuous search).

Table 2. Objectives or Fitness Functions Used. objective

description

PeakCount RTM ANND ASN

peak count after adjustment for noise and duplicates run time average nearest peakspeak neighbor distance of 25% worst peaks average signal-to-noise of 10% worst peaks

to a Leco Pegasus IV time-of-flight (TOF) mass spectrometer (Leco Corp., St. Joseph, MO)) that we used exploit a four-jet nitrogen-based cryogenic modulation system. Columns 1 and 2 were, respectively, DB-1 (30 m × 250 µm × 0.25 µm; Agilent J&W Scientific) and BPX-50 (1.5 m × 100 µm × 0.1 µm; SGE, Milton Keynes). The Windows-based ChromaTof v2.25 software was obtained from the manufacturer and ran on an IBM-compatible PC. It was employed for instrument control and raw data processing, including chromatographic deconvolution, but does not have a suitable application programming interface by which an external control program could be used to automate acquisition and processing of GC×GC-MS data. Thus, to automate this process, we are required to “mimic” manual operator input. To achieve this, a Windows macro recorder, Eventcorder (http:// www.eventcorder.com/), was used to capture and play back manual keyboard/mouse movements. As well as its own scripting environment, Eventcorder itself has its own ActiveX API and therefore playback can be controlled via any ActiveX aware programming language, such as Visual Basic, Delphi, etc. This enables playback of recorded scripts under program control with variable text (keystrokes) sent to the client program, as well as access to the full suite of features provided by the programming language. To control Eventcorder (and hence the LECO software), we (S.O’H.) developed a Microsoft VB6 Program, GCTofControl V1.8; this software also acted as the interface to the heuristic algorithm (WinParEGO) used for choosing the parameters, read in the Leco 466 Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

optimization direction maximize minimize maximize maximize

peak list exported data files (as ASCII CSV format) and calculated fitness values from these. The algorithm used, WinParEGO, is a C/C++ dynamic link library based on the command line ParEGO implementation developed by J.D.K. and ported from Linux to Windows. The ParEGO algorithm46 is a multiobjective version of the efficient global optimization (EGO) algorithm of Jones and colleagues .47 It uses a design and analysis of computer experiment (DACE)48-50 approach to model the fitness landscape(s), based on an initial “Latin hypercube” sampling of the parameter space. Subsequently, the model is used to suggest the next experiment (set of instrumentation parameter values), such that the “expected improvement” in the fitness function is maximized. The notion of expected improvement implicitly ensures that ParEGO balances exploration of new parameter combinations with exploitation and fine-tuning of parameter values that have led to good “fitnesses” in previous experiments. The DACE model is updated after each fitness evaluation. The overall arrangement is given in Figure 1. Instrument Parameters. The instrument parameters that were chosen for optimization are listed in Table 1; the oven 1 (46) Knowles, J. IEEE Trans. Evol. Comput. 2006, 10, 50-66. (47) Jones, D. R.; Schonlau, M.; Welch, W. J. J. Global Opt. 1998, 13, 455-492. (48) Sacks, J.; Welch, W.; Mitchell, T.; Wynn, H. Stat. Sci. 1989, 4, 409-435. (49) Crary, S. B. Analog Integr. Circ. Signal Proc. 2002, 32, 7-16. (50) Chen, V. C. P.; Tsui, K. L.; Barton, R. R.; Meckesheimer, M. IIE Trans. 2006, 38, 273-291. (51) Hicks, C. R.; Turner, K. V., Jr. Fundamental concepts in the design of experiments, 5th ed.; Oxford University Press: Oxford, 1999.

Figure 2. Auxiliary fitness function GMax. A GMax-bio model of an analyst’s classification of the run quality; although not used during optimization, this provided a useful aid for subsequent analysis.

start temperature and the detector voltage were fixed for the optimization runs we carried out, but could also be set from within the GCTofControl control program. The initial oven temperature was chosen, in a preoptimization experiment to ensure that metabolites of greatest volatility were detected, since too high a temperature would mean that such metabolites will elute in the solvent front (where data are not collected to increase filament and detector lifetimes). The detector voltage was set at 1700 V, initially and an equivalent sensitivity throughout the optimization experiment was ensured by determination of the S/N for mass 69 of a calibrant gas (PFTBA) on a daily basis and adjustment of the detector voltage to maintain a similar S/N and sensitivity (4 adjustments were performed). In the initial set of 50 experiments, parameters were allocated using the Latin hypercube approach, which attempts to distribute the parameter values across the parameter space (as in standard design of experiments strategies51,52). Subsequently, experimental parameters were generated using the ParEGO genetic algorithm’s DACE model. The set of 50 hypercube-generated experiments gives the ParEGO algorithm sufficient data upon which to build its initial DACE model. For experiments up to number 217, a somewhat coarser discretization of parameter ranges was used (Table 1). However, for experiments 218-246, the parameter increments were decreased to the lowest limit that would remain as acceptable inputs to the Leco software in an attempt to make the parameters approximate more nearly continuous functions. This was done because there are theoretical grounds for believing that the DACE model of the ParEGO algorithm would operate more efficiently with (more nearly) continuous values. After experiment 246, an optimal set of conditions was being approached as shown by the same parameter values being employed for consecutive experiments. The data were surveyed to describe the optimal conditions for each parameter. After optimal conditions were determined, a (52) Montgomery, D. C. Design and analysis of experiments, 5th ed.; Wiley: Chichester, 2001.

more targeted set of experiments were designed (by W.B.D.) to explore the local search space around these optimal conditions more systematically by experimentally varying one parameter while maintaining the other parameters at a constant level (experiments 250-300). Fitness (Objective) Functions. In many areas of bioanalysis, it is not possible to know a priori the nature of the compounds of interest.53 In metabolomics in particular, where it is desired to obtain information on as many (perhaps previously unknown) metabolites as possible, it is not feasible to use a targeted compound approach to analysis. Also, it is becoming important to construct experiments such that the data obtained (and the metadata54,55) may be reused at a future datesperhaps for an entirely unforeseen application. It is therefore prudent to capture as much analytical information as possible. For this reason, our primary “fitness function” or objective was chosen to be the number of peaks detected in the chromatogramsincluding appropriate filtering56 of noise and potential duplicate peaks (see below: Data Preprocessing). In a typical metabolomic experiment, it is likely that many hundreds if not thousands of samples will need to be analyzed, so the overall time taken to do each analysis becomes critical to the practicality of the experiment. The analytical run time was chosen as the next most important objective. In GC×GC, the elution time chromatogram is characterized by two time dimensions, corresponding to the retention time on (53) Smith, C. A.; Want, E. J.; O’Maille, G.; Abagyan, R.; Siuzdak, G. Anal. Chem. 2006, 78, 779-787. (54) Jenkins, H.; Hardy, N.; Beckmann, M.; Draper, J.; Smith, A. R.; Taylor, J.; Fiehn, O.; Goodacre, R.; Bino, R.; Hall, R.; Kopka, J.; Lane, G. A.; Lange, B. M.; Liu, J. R.; Mendes, P.; Nikolau, B. J.; Oliver, S. G.; Paton, N. W.; Roessner-Tunali, U.; Saito, K.; Smedsgaard, J.; Sumner, L. W.; Wang, T.; Walsh, S.; Wurtele, E. S.; Kell, D. B. Nat. Biotechnol. 2004, 22, 16011606. (55) Spasic, I.; Dunn, W. B.; Velarde, G.; Tseng, A.; Jenkins, H.; Hardy, N. W.; Oliver, S. G.; Kell, D. B. BMC Bioinformatics 2006, 7, 281. (56) Brown, M.; Dunn, W. B.; Ellis, D. I.; Goodacre, R.; Handl, J.; Knowles, J. D.; O’Hagan, S.; Spasic, I.; Kell, D. B. Metabolomics 2005, 1, 35-46.

Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

467

Figure 3. Typical GC×GC-TOF-MS 3D chromatograms from (a) experiment 7, one of the better runs during the first 50 experiments, and (b) from the final optimized set of conditions.

the first column and the retention time on the second column. A good interpretation of the resolution between two peaks, which takes into account both time dimensions, would be the diagonal distance between them on the 2D plane; the closest peak to any given peak will then be its nearest neighbor. For each peak, we calculated the nearest-neighbor distance after removal of noise and duplicate peaks. Ignoring the fact that each pair of peaks would be represented twice, we used the average of the lowest 25% nearest-neighbor distances as our third fitness function. In practice, because the interval between peaks on the first column is always larger than the interval between peaks on the second column, when there is more than one peak at a given first column elution time (as will happen most of the time), this fitness measure reduces to the time separation on the second column. 468

Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

For the final fitness function, we chose to measure the average signal-to-noise ratio of the peaks in the run. However, as nearly all peaks had very good signal-to-noise, we limited the calculation to the average of the worst 10% of peaks, so that the fitness function would reflect improvements (if any) in the worst peaks observedseither through the reduction in noise peaks or improvement in intensity of “true” peaks. These objectives are set out in Table 2. Data Preprocessing. The Leco software exports a text file containing a 2D peak list, that is 2D in the sense that chromatographic peaks are characterized by two retention time dimensions, that of the first GC column and that of the second GC column. Although each peak can be expanded into a full mass spectrum, and the data are therefore multidimensional, we only utilized the

Figure 4. Improvement in peak number during the closed-loop optimization of GC×GC-TOF-MS measurements of human serum. The size of the symbol encodes the run time while the color (red low, blue high) encodes the nearest-neighbor distance defined above. The last 6 experiments in green represent validation replicates using the final chosen conditions. (a) Peaks vs experiment number. (b) Peaks vs run time.

peak area of the quantification ion identified by the peak deconvolution software for our analysis. (Note: due to the nature of the deconvolution algorithm, there is no guarantee that, for a given peak, the same quantification ion will be chosen from run to run, so this may introduce some bias; however, this is unavoidable as we have little or no control over deconvolution parameters). Due

to the Leco peak detection algorithm and the way that GC “slices” from the first column are trapped and fed into the second column, there is a high potential for the occurrence of duplicate or multiplet chromatographic peaks to be identifiedsi.e., multiple peaks in the output file are in fact only one chemical entity. The match required to combine was set at 500, a reasonable value to ensure modulated Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

469

Table 3. Auxiliary Fitness Functions. auxiliary fitness function

description

IQR10

interquartile range of second dimension elution time, after applying a threshold of 10 on signal-to-noise IQR10 as described above divided by the width of the second dimension time window overlap of second dimension elution time peak-count histogram with perfect rectangular distribution of same total area overlap of second dimension elution time peak-area histogram with perfect rectangular distribution of same total area the virtual analyst: A GMax-Bio model based on an analyst’s qualitative judgment of a subset of runs

IQR10/SDTW LBc LBa GMax

Table 4. Set of Final Optimized Conditions after 300 Experiments. variable parameters

units

sample volume corrected column flow split ratio inlet temperature oven 1 start hold time oven 1 ramp rate oven 1 final temperature final temperature duration oven 2 start temperature oven 2 start hold time oven 2 ramp rate oven 2 final temperature final temperature duration transfer line temperature modulator temperature offset second dimension time hot pulse time acquisition rate ion source temperature

µL mL/min °C min °C/min °C min °C min °C/min °C min °C °C s s Hz °C

optimized value 3 1 01:15 260 5 5 290 10 57 5 5 290 10 220 25 6 0.3 130 240

peaks of the same metabolite are combined. Another problem is that the position of the initial solvent front changes with instrument parameters (which the optimization algorithm alters from run to run), so that using a fixed solvent delay runs the risk of either losing genuine peaks if the solvent delay is too long or passing unwanted solvent peaks if the solvent delay is too short in comparison with the position of the solvent front. Both of these effects would lead to an incorrect estimate in the peak number fitness, and therefore, simple preprocessing steps were applied prior to calculating fitness. Initial noise removal involves removing all peaks from the peak list with a peak area of 750), showing the major need to identify these metabolite peaks by running authentic standards and thereby including more known (or estimated) metabolites in mass spectral libraries. It should be noted that many metabolites produce multiple derivatization (69) Kochhar, S.; Jacobs, D. M.; Ramadan, Z.; Berruex, F.; Fuerhoz, A.; Fay, L. B. Anal. Biochem. 2006, 352, 274-281. (70) Lenz, E. M.; Bright, J.; Wilson, I. D.; Hughes, A.; Morrisson, J.; Lindberg, H.; Lockton, A. J. Pharm. Biomed. Anal. 2004, 36, 841-849. (71) Kenny, L. C.; Dunn, W. B.; Ellis, D. I.; Myers, J.; Baker, P. N.; The GOPEC Consortium; Kell, D. B. Metabolomics 2005, 1, 227-234; online DOI: 210.1007/s11306-11005.-10003-11301. (72) Underwood, B. R.; Broadhurst, D.; Dunn, W. B.; Ellis, D. I.; Michell, A. W.; Vacher, C.; Mosedale, D. B.; Kell, D. B.; Barker, R.; Grainger, D. J.; Rubinsztein, D. C. Brain 2006, 129, 877-886. (73) Romero, P.; Wagg, J.; Green, M. L.; Kaiser, D.; Krummenacker, M.; Karp, P. D. Genome Biol. 2005, 6, R2. (74) Williams, R.; Lenz, E. M.; Wilson, A. J.; Granger, J.; Wilson, I. D.; Major, H.; Stumpf, C.; Plumb, R. Mol. Biosyst. 2006, 2, 174-183. (75) Brereton, R. G.; Dunkerley, S. Analyst 1999, 124, 705-711. (76) Demir, C.; Hindmarch, P.; Brereton, R. G. Analyst 2000, 125, 287-292. (77) Woodward, A. M.; Rowland, J. J.; Kell, D. B. Analyst 2004, 129, 542-552. (78) Sinha, A. E.; Hope, J. L.; Prazen, B. J.; Fraga, C. G.; Nilsson, E. J.; Synovec, R. E. J. Chromatogr., A 2004, 1056, 145-154. (79) Smilde, A.; Bro, R.; Geladi, P. Multi-way analysis: applications in the chemical sciences; Wiley: New York, 2004. (80) A, J.; Trygg, J.; Gullberg, J.; Johansson, A. I.; Jonsson, P.; Antti, H.; Marklund, S. L.; Moritz, T. Anal. Chem. 2005, 77, 8086-8094. (81) Katajamaa, M.; Miettinen, J.; Oresˇicˇ, M. Bioinformatics 2006, 22, 634636.

476

Analytical Chemistry, Vol. 79, No. 2, January 15, 2007

products, and therefore, in this study more than 350 peaks have been assigned an identification, and it is this that is equivalent to 188 metabolites. Current work is ongoing to identify metabolites definitively in this way by the compilation of a mass spectral/ retention index library for the optimized set of conditions reported, including those for metabolites not currently available in academic or commercially available mass spectral libraries. Based on the heuristic analysis of the separations landscape in this system, it does not appear likely that major improvements in the number of peaks will now come from improving the chromatographic separations per se. However, we recognize that much greater resolution in the mass spectral dimension is possible, including the use of exact mass techniques,74 and this is evidently an important strategy both for further improving the number of metabolites that may be detected and for assisting their identification. Another approach that we are pursuing is to seek to identify in the 2D GC×GC-TOF-MS data all the metabolites that we would expect from the known biochemistry to be present in human serum. If the deconvolution problem is reduced to deconvolving known coeluting substances from each other, this will permit the application of different and more powerful chemometric techniques (cf. refs 53 and 75-81). In conclusion, by using an advanced, closed-loop optimization method, we have demonstrated a substantial improvement in the number of human serum metabolites that may be detected reliably using GC×GC-TOF-MS. Understanding their nature and distribution between individuals under different conditions is now possible, where the existence of a suitable data model and database55 will make this task much easier. ACKNOWLEDGMENT We thank the BBSRC and MRC, partly under the terms of the LINK scheme in Applied Genomics (the HUSERMET project), for financial support, as well as the EPSRC, RSC, and BHF. We also thank AstraZeneca and GlaxoSmithKline for support of the HUSERMET project. We thank Nick Bukowski (Anatune), Roy Goodacre (Manchester), John Haselden (GSK), Andy Nicholls (GSK), and Ian Wilson (AZ) for useful discussions. S.O’H., W.B.D., and J.D.K. contributed equally to this work. Received for review August 3, 2006. Accepted October 25, 2006. AC061443+